Code Sandbox

Code Sandbox

Routes code execution and AI coding tasks to the right tool.

Routing Table

SandboxFusion — Safe Code Execution

SandboxFusion (ByteDance) provides isolated, multi-language code execution for LLM benchmarking and untrusted code evaluation.

Quick Start (Docker)

# Pull and run SandboxFusion server
docker pull bytedance/sandbox-fusion:latest
docker run -d -p 8080:8080 --privileged bytedance/sandbox-fusion:latest

# Or with docker-compose
git clone https://github.com/bytedance/SandboxFusion.git
cd SandboxFusion
docker compose up -d

Execute Code via API

import requests

SANDBOX_URL = "http://localhost:8080"

def execute_code(code: str, language: str = "python") -> dict:
    response = requests.post(f"{SANDBOX_URL}/run_code", json={
        "code": code,
        "language": language,
        "timeout": 10,  # seconds
    })
    return response.json()

# Run Python
result = execute_code("print('Hello, world!')", "python")
print(result["stdout"])  # Hello, world!
print(result["stderr"])  # (empty)
print(result["exit_code"])  # 0

# Run JavaScript
result = execute_code("console.log(2 + 2)", "javascript")

# Run with test cases (for LLM eval)
response = requests.post(f"{SANDBOX_URL}/run_code", json={
    "code": "def add(a, b): return a + b",
    "language": "python",
    "test_cases": [
        {"input": "add(1, 2)", "expected": "3"},
        {"input": "add(-1, 1)", "expected": "0"}
    ]
})

Supported Languages

Python, JavaScript, TypeScript, Java, C, C++, Go, Rust, PHP, Ruby, Bash, R, Scala, Kotlin

LLM Benchmark Mode

# Evaluate LLM-generated code against test suites
response = requests.post(f"{SANDBOX_URL}/submit", json={
    "dataset": "humaneval",
    "problem_id": "HumanEval/0",
    "solution": llm_generated_code,
    "language": "python"
})
print(response.json()["passed"])  # True/False

Security Properties

• Each execution runs in isolated container
• No network access by default
• Resource limits (CPU, memory, time)
• No filesystem persistence between runs

Aider — AI Pair Programming

Aider is a terminal-based AI coding assistant that works with your existing git repo.

Installation

pip install aider-chat
# or
pipx install aider-chat

Basic Usage

# Start Aider in your repo
cd /path/to/your/project
aider

# With specific model
aider --model gpt-4o
aider --model claude-opus-4-5-20251001  # Use with Claude

# Add specific files to context
aider src/main.py tests/test_main.py

# Pass instruction directly
aider --message "Add input validation to the user registration function" src/auth.py

Common Workflows

# Fix a bug
aider --message "Fix the KeyError in process_data() when key is missing" src/processor.py

# Add a feature
aider --message "Add pagination support to the /users endpoint" src/api/users.py

# Write tests
aider --message "Write pytest unit tests for all functions in utils.py" src/utils.py tests/test_utils.py

# Refactor
aider --message "Refactor this class to use composition instead of inheritance" src/models.py

# Review and explain
aider --message "Explain what this function does and suggest improvements" src/complex_algo.py

Key Flags

--model MODEL          # LLM to use (gpt-4o, claude-opus-4-5-20251001, etc.)
--no-git               # Don't use git (disable auto-commit)
--yes                  # Auto-confirm all prompts
--no-auto-commits      # Make changes but don't auto-commit
--read FILE            # Add file as read-only context (no edits)
--message "..."        # Send message without interactive mode
--edit-format diff     # Use diff format for edits (default: udiff)
--cache-prompts        # Cache API calls (reduce cost on retries)

Environment Setup

# Set API key (pick one)
export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...

# Use Claude models
aider --model claude-opus-4-6
aider --model claude-sonnet-4-6

# Use local model via Ollama
aider --model ollama/codellama:34b

.aider.conf.yml (project config)

# .aider.conf.yml in project root
model: claude-sonnet-4-6
no-auto-commits: true
read:
  - README.md
  - ARCHITECTURE.md

Decision Guide

"Run code from an LLM safely" → SandboxFusion "Evaluate if LLM-generated code passes tests" → SandboxFusion "Add a feature to my existing repo" → Aider "Fix a bug with AI help" → Aider "Write tests for existing code" → Aider "Multi-language code execution in CI" → SandboxFusion API

Integration: Aider + Claude Code

When working inside Claude Code, use Aider for:

• Bulk edits across many files (Aider's whole-repo context)
• Applying a diff from Claude Code suggestions
• Iterative coding with automatic git commits

# Apply Claude's suggestion using Aider
aider --message "$(cat suggestion.txt)" --yes src/target.py

Environment Variables

OPENAI_API_KEY=          # For Aider with GPT-4
ANTHROPIC_API_KEY=       # For Aider with Claude
SANDBOX_FUSION_URL=      # SandboxFusion server URL (default: http://localhost:8080)